Raw waveform speaker verification for supervised and self-supervised learning

Speaker verification models that directly operate upon raw waveforms are receiving growing attention. However, their performances are less competitive than the state-of-the-art handcrafted feature-based counterparts, demonstrating equal error rates under 1% on the benchmark VoxCeleb1 evaluation protocol. In addition, they have yet not been explored with self-supervised learning frameworks. This paper proposes a new raw waveform speaker verification model that incorporates techniques proven effective for speaker verification, including the Res2Net backbone module and the aggregation method considering both context and channels. Under the best performing configuration, the model shows an equal error rate of 0.89%, competitive with state-of-the-art models. We also explore the proposed model with a self-supervised learning framework and show the state-of-the-art performance in this line of research. Finally, we show that leveraging the model trained with self-supervision successfully serves as a pre-trained model under the semi-supervised scenario where it is assumed that only a limited amount of data has a ground truth label and a bigger data has no label.